There's a GitHub repo called awesome-agent-skills with 29.4k stars that indexes free, ready-made skills for Claude Code, Cursor, Codex and friends. It's genuinely useful. It's also badly organised for finding anything — it's sorted by who published, not by what you need. This guide fixes that.
First, the number — because three sources disagree
The repo's description says 1000+. The badge at the top of its own README says 1497+. The official directory site says 648.
They're measuring different things. 648 is the count of official skills from 53 named dev teams, and it's the figure with a dated timestamp on it. The larger number appears to include community-submitted skills, which aren't vetted the same way. 648 is the number I'd stand behind, and it's the one used throughout this guide.
Worth knowing before you repeat a figure you saw in a video.
What's in this guide
1. What an agent skill actually is
A skill is a folder with a SKILL.md file in it. That file describes a task in detail — the steps, the gotchas, the tools to use — and your AI assistant loads it only when the task comes up.
The distinction that matters:
- A prompt is something you type every time and it's gone when the chat ends.
- An MCP server connects your assistant to an external service — it adds capability.
- A skill adds procedure. It's the written-down knowledge of how a task should be done, and it persists.
So a skill published by Stripe's team is Stripe's own engineers writing down how to do a thing properly with Stripe. That's the actual pitch here — not "more prompts," but vendor documentation your assistant reads automatically.
2. The install command
The repo never tells you this. It's on the directory site instead, and it's one line:
npx skills add https://github.com/anthropics/skills --skill skill-creatorSwap the repo URL and the skill name for whichever one you want. The pattern is always npx skills add <repo-url> --skill <skill-name>.
If that fails or you'd rather not run it, there's a second route that works surprisingly well: paste the skill's GitHub link into your assistant and ask it to install the skill. Every skill page gives you a direct link for exactly this.
Where they work: Claude Code, Codex, Cursor, GitHub Copilot, OpenCode and Antigravity are all listed as supported, with Gemini CLI and Windsurf named in the README too. Install paths differ per tool — if a skill doesn't trigger, the path is the first thing to check.
3. All 9 categories, by size
These are the official categories and counts, read off the directory on 11 August 2026. They add up to exactly 648.
| Category | Skills | What's in it |
|---|---|---|
| infrastructure | 147 | Deploying, hosting, databases, cloud |
| development | 105 | Language and framework specifics |
| ai-tools | 90 | Models, media generation, agent building |
| testing | 68 | Test frameworks and QA automation |
| security | 64 | Auditing, auth, vulnerability review |
| workflows | 61 | Marketing, content, ops, productivity |
| data | 56 | Analytics, warehouses, scraping |
| design | 37 | Figma, UI, visual and video work |
| docs | 20 | Writing and maintaining documentation |
Two things jump out. Infrastructure and development are more than a third of everything — this collection is built by and for engineers shipping production software. And docs is the smallest category by a wide margin, at 20 skills, which is a fair summary of how the industry treats documentation.
infrastructure — 147
The biggest category, and the most "official." Cloudflare, Netlify, Vercel, Firebase, Supabase, Neon, Redis, MongoDB, Google Cloud, HashiCorp and Red Hat all publish here. If you deploy anything, your host probably has a skill that knows its own quirks better than a general model does.
Example: netlify/netlify-deploy — "Automated deployment workflow for Netlify sites."
development — 105
Framework and language specifics. Angular, Expo, Flutter, Apollo GraphQL, Better Auth, Auth0 and WordPress are all here. Microsoft alone publishes 133 skills across .NET, Java, Python, Rust, TypeScript and core — which tells you the shape of the whole collection.
This is the category that solves the version-drift problem. A model trained last year will confidently write last year's API; a skill maintained by the framework team won't.
ai-tools — 90
The most interesting category if you're not a full-time engineer. Anthropic, OpenAI, Hugging Face, Replicate, fal.ai, Brave and Venice.ai publish here.
Examples, all verbatim from their listings:
anthropics/skill-creator— "A skill for building, testing, and refining other skills."anthropics/mcp-builder— "A development guide for building MCP servers that connect LLMs to external APIs and services."replicate/replicate— "Discover, compare, and run AI models using Replicate's API."fal-ai-community/fal-lip-sync— "Create talking head videos and lip sync audio to video."fal-ai-community/fal-video-edit— "Edit existing videos using AI — remix style, upscale, remove background, add audio."fal-ai-community/fal-train— "Train custom AI models (LoRA) on fal.ai for personalized image generation."brave/bx— "A CLI tool for web search built for AI agents."
testing — 68
Test frameworks and QA automation. Worth knowing that TestMu AI publishes the largest section here and is also a paid sponsor of the repo — see the caveats at the end.
Some entries carry their own deprecation notices, which is a good sign for honesty and a bad sign for currency: testmu-ai/protractor-skill is marked "deprecated; prefer Playwright/Cypress," and testmu-ai/lettuce-skill as "legacy; prefer Behave." Read the label before installing.
security — 64
Trail of Bits — a well-regarded security firm — publishes here, alongside Sentry, Auth0, Coinbase and Binance. This is the category where "written by people who do this professionally" carries the most weight, because security advice from a general model is exactly the kind you shouldn't trust.
workflows — 61
The category most likely to be useful if you're not an engineer, and the one buried deepest in the repo.
Corey Haines publishes around 31 marketing skills covering the full stack — SEO audits, copywriting, email sequences, pricing strategy, landing page CRO, referral programmes, paid ads. Kim Barrett publishes direct-response advertising skills including headline generation, objection handling and Schwartz-style awareness mapping.
Examples:
typefully/typefully— "Create, schedule, and publish social media content across X, LinkedIn, Threads, Bluesky, and Mastodon."coreyhaines31/email-sequence— "Build email sequences, drip campaigns, and lifecycle email flows."coreyhaines31/page-cro— "Improve conversion rates on any marketing page including homepages and landing pages."realkimbarrett/headline-matrix— "Generate high-performing headline variations across different angles."googleworkspace/gws-sheets— "Read and write Google Sheets spreadsheets."
data — 56
ClickHouse, DuckDB, Tinybird, MongoDB, Redis and Datadog Labs. Also Firecrawl and Browserbase, which handle scraping and browser automation — the two things people most often try to build badly from scratch.
design — 37
Figma publishes its own official set, and OpenAI publishes a parallel set for the Figma MCP server. Google Labs contributes Stitch skills, and there's GSAP for animation and Remotion for video.
Examples:
figma/figma-implement-design— "Translate Figma designs into production-ready application code with 1:1 fidelity."remotion-dev/remotion— "Programmatic video creation with React."anthropics/canvas-design— "Design visual art in PNG and PDF formats."anthropics/algorithmic-art— "Create generative art using p5.js with seeded randomness."openai/slides— "Create and edit .pptx presentation decks with PptxGenJS."
docs — 20
The smallest category. Notion and Sanity publish here, and Addy Osmani contributes a web quality set. Small, but the skills are the kind you install once and forget about.
4. The ones worth installing first
648 is paralysing. Install three, not thirty.
If you build things: the skill for whatever you deploy on — Netlify, Vercel, Cloudflare, Supabase. Highest hit rate of anything here, because deployment is where models are most confidently wrong about current APIs.
If you make content: typefully/typefully for scheduling across platforms, and Corey Haines's page-cro and copywriting. The marketing set is the least-known good thing in the whole collection.
If you make video: the fal.ai set — lip-sync, video-edit, restore, train. And remotion-dev/remotion if you're comfortable in React.
If you want to understand the system itself: anthropics/skill-creator. It builds, tests and refines other skills, including running comparison evals with and without a skill active. Once you've used it, you'll write your own instead of hunting for someone else's.
5. What the repo doesn't tell you
Three things worth knowing
It's an index, not a library. The repo contains almost no skill code. Nearly every entry links out to a directory site or the publisher's own repo. Cloning it gets you a list, not 648 skills.
There are paid sponsors inside the list. TestMu AI, Zero and Ego Lite are sponsors, and two more products are promoted inside the table of contents itself. TestMu AI also holds the largest non-Microsoft skill section. The repo's tagline is "hand-picked, not AI-slop generated" — that's likely true of the official sections, and it sits awkwardly beside paid placement in the same list. Judge the sponsored entries on their merits, not their position.
"Official" and "community" are not the same tier. The 648 figure covers skills from 53 named dev teams. The larger 1497+ figure includes community submissions. A skill from Stripe's engineers and a skill from an anonymous contributor both look identical in a list. Check who published before you let one run.
None of this makes it a bad resource — it's the best index of its kind, it's MIT licensed, and having Anthropic, Stripe, Figma and Trail of Bits writing procedural documentation for your assistant is a real shift in how this stuff works.
Just install deliberately. A skill is instructions your assistant will follow without asking you again, which is exactly why you should know who wrote it.
Method note: Counts and category names were read directly off the official directory on 11 August 2026, not from the repo README — which is how the 648 / 1000+ / 1497+ discrepancy turned up. Every skill description quoted here is verbatim from its own listing. These numbers move weekly; re-check before you rely on them.
Want the next one first?
The Input Daily — a 5-minute AI brief every weekday. 3 stories that matter, 1 tool worth trying, 1 prompt to steal.
Free. Every morning. Unsubscribe in one click. Privacy Policy.
You're all set.
The Input Daily lands in your inbox tomorrow morning. While you're here — the rest of the guides are free too.
Browse all free guides →